API to get model native type

[marcobambini]

I am using llama.cpp from C to generate embedding, and I can get the embedding dimension using

int dimension = llama_model_n_embd(llama_get_model(ctx));

Internally llama.cpp knows the model native type (for example, for Qwen3-Embedding-0.6B-f16.gguf is float16).

Is there a way to get that type using the public APIs?

[marcobambini]

How safe is it to use the value from the key general.file_type?

[ggerganov]

In most cases yes, though this key is optional. Meaning some GGUFs might not have it.

There isn't a specific API to do that. You could probably analyze the types of all tensors and deduce the overall quantization type. For example, if the majority of tensors are type Q8_0, you could report that.